[Qemu-devel] QEMU live migration RARP issue

[Robin Geuze <robing@transip.nl>]

Hey guys,

A few days ago we noticed that some of our VPSes lost network connection 
after a live migration using QEMU/KVM. We started investigating and 
noticed that the problem was at least present for guests running 
FreeBSD, Debian 7 (Linux 3.2.81-1) and some flavors of CentOS. At least 
Debian 8 (Linux 3.16.7-ckt25-2) was not affected. Everything is using 
the virtio drivers. We also noted that the problem was not present in 
QEMU 2.4.0, but was present in QEMU 2.6.0. The configs between both 
versions of QEMU are identical. We use libvirt 1.3.5 to manage QEMU/KVM 
(same version on both qemu versions).

After some false starts we managed to reliably reproduce the issue with 
a Debian 7 VM. What we noticed was that Debian 8 send out gratuitous ARP 
replies. Debian 7 did not appear to send those. After some further 
digging we found out that on QEMU 2.4.0 Debian 7 actually send 
gratuitous RARP packets after a live migration instead of normal ARP 
packets and that those where not present when migrating on QEMU 2.6.0. 
We then started narrowing down the responsible commits using git bisect 
and eventually ended up with the last good commit being 
d62241eb6da9bd2517f07b3219ba4208b90b4e0d and the last bad commit we 
tested was 7ef7bc8586fb0d41742a896b532c7afa2bbb7f84. This is a range of 
6 commits introducing the NetFilter functionality. We couldn't really 
pinpoint it more accurately so we decided to email you guys, to see if 
you can figure out whats up.

Regards,

Robin Geuze

TransIP BV